CN104331699B - A kind of method that three-dimensional point cloud planarization fast search compares - Google Patents

Abstract

The invention discloses a method for quick search and comparison of three-dimensional point cloud planarization. The method first obtains point cloud data of an object, performs smoothing and streamlining processing on it, and then selects a two-dimensional view according to requirements, and finds the boundary of the view. , and then based on this boundary, the image is divided into grids according to the requirements, and each grid of the segmented image is traversed, and the grid is marked according to the point cloud density in each grid, and then used The obtained marking results make an approximate binary image, which has feature points that can reflect the point cloud data, and then use the scale-invariant feature transformation matching algorithm to compare the approximate binary image with the image processed by the same method in the standard library Perform feature point comparison, and use traversal to find a set of data with the most matching feature points. This method has the characteristics of high precision, fast speed, and high flexibility, and is suitable for various occasions that need to establish a standard library and quickly register point cloud data with images in the library.

Description

A method for rapid search and comparison of 3D point cloud planarization

technical field

The invention relates to a method for three-dimensional point cloud processing and three-dimensional point cloud matching, which belongs to the field of computer vision and pattern recognition, and in particular to a method for fast search and comparison of three-dimensional point cloud planarization.

Background technique

3D point cloud data refers to the spatially distributed mass point collection of the morphological structure of an object obtained through 3D digital technology. High-precision point cloud data can better represent the three-dimensional shape characteristics of the measured object. It is used in the development of molds and products such as automobiles, hardware appliances, aviation, ceramics, rapid prototyping of antiques, handicrafts, sculptures, portrait products, and mechanical shape design. It has important applications in fields such as medical plastic surgery, human body shape making, human body shape measurement and plant morphology acquisition.

Most of the traditional 3D graphic matching is to directly compare the point cloud data to be matched with the standard point cloud data, and perform similarity or identity analysis after geometric transformations such as translation, rotation, and scaling. This technology is widely used in various fields such as medicine, civil engineering, and reverse engineering in industry. High precision and good matching effect, but the traditional matching method has high space-time complexity, and when the model has too much point cloud data, its calculation time will be greatly increased. Although this method is suitable for precise matching, it is only suitable for pairwise Matching is not suitable for the occasion of matching several point cloud data in the same standard library, nor is it suitable for the occasion that requires quick identification.

Contents of the invention

In view of this, the present invention proposes a method for rapid search and comparison of three-dimensional point cloud planarization for the shortcomings of existing methods. This method has the characteristics of high precision, good speed, and high flexibility, and is suitable for various Occasions that need to be quickly registered with the point cloud data in the standard library.

The technical solution of the present invention is: preprocessing the 3D point cloud, converting the 3D point cloud image into a 2D graphic, performing grid segmentation on the 2D graphic, and generating an approximate binary value for matching based on the segmented graphic The image is matched with the image in the standard library according to the approximate binary image, so as to find out a set of point cloud data in the standard library that is closest to the three-dimensional point cloud data of the object to be measured. Specifically include the following steps:

Step 1: Obtain the 3D point cloud of the object to be measured, and then use the bilateral filter denoising algorithm to smooth the point cloud data, and then use the random sampling method to simplify the data of the smoothed point cloud data, and finally the streamlined point cloud data The point cloud data is subjected to two-dimensional transformation, and the selected view (referring to the front view and side view) must be consistent with the view used when building the library in the standard library, and the two-dimensional point cloud image is generated after dimensionality reduction;

Step 2: Find the four boundary points of the two-dimensional point cloud image through quick sorting, then generate the two-dimensional point cloud image boundary according to these four boundary points, and perform grid segmentation on the image, and then retrieve each segmented The grid, according to the density difference of the point cloud in different grids, makes corresponding marks, and fills the corresponding colors according to the marks, and generates an approximate binary image corresponding to the gradient;

Step 3: Use the scale-invariant feature transformation matching algorithm to compare the obtained image with the approximate binary image in the standard library. After all the comparisons are completed, use sequential search to compare each group of images in the standard library Match the feature points of the group, find the group with the most corresponding feature points, and complete the comparison.

Further, the two-dimensional transformation in the step 1 refers to: according to the need to select one of the three views for generating a three-dimensional image according to the comparison, the adopted view must be the same view as the image in the standard library to ensure accuracy .

Further, the specific steps of the bilateral filtering denoising algorithm in the first step are: 3.1 establishing K-neighborhood; 3.2 normal vector estimation; 3.3 defining the viewing plane; 3.4 introducing a bilateral filtering operator to obtain smoothed coordinates.

Further, the random sampling method in step 1 refers to: establish a function so that the random numbers it generates can just include all point clouds, and then generate a continuous set of random numbers, and then find the other random numbers from the original point cloud. Corresponding points are eliminated until the total points meet the established requirements.

Further, the boundary point in the step 2 specifically refers to the farthest point in the image, for example, in the front view, the quick sorting method is used to find the two points with the smallest and largest X values, and the points with the smallest and largest Y values. Two points, a total of four coordinate points are its boundary points.

Further, the grid division in step 2 refers to dividing the image into n×n (n∈R, n>0) grid blocks based on the boundary determined in this step.

Further, the approximate binary image in the step 2 refers to: because the color of the image made is related to the density mark marked in step 2, the image made is between the grayscale image and the binary image, for convenience Expression, the two-dimensional image used for comparison is collectively called an approximate binary image, which can generate more feature points through the marking density to improve the accuracy of the method.

Further, generating an approximate binary image corresponding to the gradient in step 2 specifically includes the following steps: traversing the point cloud in each grid, marking each grid according to the number of point clouds in the grid, According to the obtained corresponding marks, fill the corresponding grid with the corresponding color in turn, and generate a pixel map with a size of n×n (n∈R, n>0). For the convenience of calculation and to ensure that enough feature points can be generated , take two colors of black and gray to represent different densities, and the grid without point cloud is represented by white.

Further, the scale-invariant feature transformation matching algorithm in the step 3 specifically includes the following steps: 9.1 Establishing image scale space; 9.2 Detecting key points; 9.3 Allocation of key point directions; 9.4 Feature point descriptors; 9.5 Using exhaustive method, Compare the feature points of the two images, and count the number of matching feature points for search and comparison.

Compared with the prior art, the present invention has the following advantages: (1) The time and space complexity is low, and the rapidity of search and comparison is improved under the condition of ensuring high precision. (2) It is suitable for various occasions that require quick comparison and identification. It is not only suitable for pairwise matching, but also can establish a standard library by itself and compare it with the data in the standard library. (3) Strong anti-interference ability. By modifying the key drawing parameters (the mark of the point cloud density in the grid), the comparison accuracy can still be guaranteed even when the smoothing effect is not good. (4) The focus of this algorithm is recognition, so when building a standard library, you can scan only one side with the most obvious features of the object, or you can scan all of them. Different methods are used according to different needs, which has extremely high flexibility and can save a lot of storage. Space, improve the comparison speed.

Description of drawings

In order to make the purpose, technical scheme and beneficial effect of the present invention clearer, the present invention provides the following drawings for illustration:

Fig. 1 is the flow chart of a kind of three-dimensional point cloud planarization fast search comparison method of the present invention

Fig. 2 is a schematic diagram of grid segmentation according to the present invention

Fig. 3 is a schematic diagram of an approximate binary image according to the present invention

Figure 4 is the rendering of the point cloud data using the Stanford Bunny rabbit

detailed description

The preferred embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

Fig. 1 is the flow chart of method of the present invention, and this method comprises the following steps:

Step 1: Decide to scan a single field of view or multiple fields of view as needed.

Step 2: Scan the object to be compared with a 3D measurement system to obtain a 3D point cloud of the object to be compared.

Step 3: If multiple fields of view need to be scanned, stitch all single-field 3D point clouds into the same measurement coordinate system. The stitching methods used include: mechanical arm-assisted stitching method, pasting marker points, etc., which are general methods for 3D point cloud stitching.

Step 4: Smooth the obtained point cloud data c={p ₁ ,p ₂ ,...,p _n }p _i ∈ R ³ with a bilateral filter denoising algorithm.

Step 5: Streamline the smoothed 3D point cloud data according to the random sampling method, and retain 40% of the point cloud data.

Step 6: Complete the preprocessed point cloud data c", and transform it into a front view or a top view as required, in this case into a front view. Remove the Z-axis part of the point cloud data in c" and keep it as (X,Y ), and the two-dimensional point cloud data is denoted as c ^* .

Step 7: Find four boundary points by traversing c ^* , the rightmost point of the image is the maximum value of X The leftmost point of the image is the minimum value point of X The uppermost side of the image is the maximum point of Y The lowermost side of the image is the minimum point of Y

Step Eight: Convert the image to Four points are the boundary, as shown in Figure 2, divided into n×n (n ∈ R, n>0) grids, each grid size is long width in Represent a point The value of X in , and so on.

Step 9: Traverse the point cloud in each grid, and mark each grid according to the point cloud density, as shown in Figure 2. When the number of point clouds is less than X ₁ , the mark is 0, when the point cloud When the number of cloud is greater than X ₁ and less than X ₂ , it is marked as 1, when the number of point cloud is greater than X ₂ , it is marked as 2 until X _i . The sizes and numbers of X ₁ , X ₂ to X _i are set according to the actual situation. In this example, only X ₁ and X ₂ are set.

Step 10: Make an approximate binary image based on the obtained grid and corresponding marks. In this example, the grid marked 0 is painted white, the grid marked 1 is drawn black, and the grid marked 2 is drawn gray , the color can be adjusted by yourself, and the drawing effect is shown in Figure 3 and Figure 4.

Step 11: Use the scale-invariant feature transformation matching algorithm to compare the obtained image with the approximate binary image in the standard library. After all the comparisons are completed, use sequential search to compare each Match the feature points of a group, find the group with the most corresponding feature points, and complete the comparison.

Further, the bilateral filtering denoising algorithm adopted in step 4 specifically includes the following steps:

4.1: Establish K-Neighborhood

Any measuring point p in the point cloud data c satisfies p∈c, then the k points closest to the measuring point p are called the K-neighborhood of p, denoted as N(p). Where k is 25.

4.2: Normal Vector Estimation

The N(p) obtained in the previous step is used to construct a plane within N(p) by the least square method, which is called the tangent plane of point p on the neighborhood N(p), denoted as T(p). The unit normal vector of T(p) is the unit normal vector at point p

4.3: Defining the view plane

Decompose the space R3 into a direct sum of two subspaces: Among them, N is the one-dimensional space along the normal vector direction of the neighborhood at point p, and S2 is the two-dimensional tangent plane space passing through point p. In the local area, S2 is defined as the viewing plane, the projection of the neighborhood point on the S2 plane is the pixel position, and the distance from the domain point to the projected point is defined as the pixel value, which is similar to image processing.

4.4: Bilateral filter operator

Introduce bilateral filter operator

Among them, N(p) is the neighborhood of p, p _i ∈ N(p). p' is the projection of p on S2. Here, the direct three-dimensional space distance is not directly used, but the distance on the projection plane is used. is the normal vector of point p, is the normal vector of _pi . W _C and W _s are Gaussian kernel functions with σ _c and σ _s as the standard deviation respectively, σ _c controls the degree of smoothness and σ _s controls the degree of feature preservation, Wc is the weight of the space domain; Ws is the weight of the feature domain. d is the distance adjusted in the direction of the normal vector, according to the formula Get the coordinate c^ after smoothing.

Further, the scale-invariant feature transformation matching algorithm adopted in step eleven specifically includes the following steps:

11.1: Establish an image scale space (or Gaussian pyramid), and detect extreme points. The points mentioned here and below are all pixels in the image.

This algorithm uses the Gaussian function to establish the scale space, and the formula of the Gaussian function is:

The above formula G(x, y, σ) is a scale-variable Gaussian function.

The scale space of an image, L(x,y,σ) is defined as the convolution of a scale-variable Gaussian function G(x,y,σ) with the original image I(x,y)

L(x,y,σ)=G(x,y,σ)*I(x,y) (3)

The scale space is represented by a Gaussian pyramid when it is implemented. The pyramid model of an image refers to a tower model that continuously downsamples the original image to obtain a series of images of different sizes, from large to small, and from bottom to top. The original image is the first layer of the pyramid, and the new image obtained by downsampling each time is a layer of the pyramid (one image for each layer), and each pyramid has n layers in total. The number of layers of the pyramid is determined according to the original size of the image and the size of the image on the top of the tower. The calculation formula is as follows:

n=log ₂ {min(M,N)}-t,t∈[0,log ₂ {min(M,N)}] (4)

Among them, M and N are the size of the original image, and t is the logarithm of the minimum dimension of the tower top image.

After the scale space is established in the scale space, in order to find stable feature points, the Gaussian difference method is used to detect those extreme points in the local position, that is, the images in two adjacent scales are subtracted, that is, the formula defined as:

D(x,y,σ)=(G(x,y,kσ)-G(x,y,σ))*I(x,y) (5)

=L(x,y,kσ)-L(x,y,σ)

11.2: Detecting Keypoints

In order to find the extreme points of the scale space, each sampling point is compared with all its neighbors to see whether it is larger or smaller than its neighbors in the image domain and scale domain. Because it needs to be compared with adjacent scales, only extreme points of two scales can be detected in a set of Gaussian difference images, while extreme point detection of other scales needs to be performed in the upper layer of Gaussian difference images of the image pyramid , so as to sequentially complete the detection of extreme values of different scales in the difference of Gaussian images of different layers in the image pyramid.

11.3: Assignment of Keypoint Orientation

In order to make the descriptor invariant to rotation, it is necessary to use the local features of the image to assign a direction to each feature point. Using the characteristics of the gradient and direction distribution of the pixels in the neighborhood of the key point, the gradient modulus and direction can be obtained as follows:

The scale is the scale of each feature point.

Sampling in the neighborhood window centered on the key point, and use the histogram to count the gradient direction of the neighborhood pixels. The gradient histogram ranges from 0 to 360 degrees, with one direction every 10 degrees, and a total of 36 directions.

The peak of the histogram represents the main direction of the neighborhood gradient at the key point, which is the direction of the key point.

11.4: Feature Point Descriptor

After the calculation of the above steps, three pieces of information have been given to each feature point: position, scale and direction. After that, a descriptor can be established for each feature point, and finally a feature vector is formed. At this time, the feature vector already has scale invariance and rotation invariance.

11.5: Use the exhaustive method to compare the feature points in the two pictures, take a certain feature point in the image to be tested, find out the first two feature points with the closest Euclidean distance in the standard image, and use the Among the feature points, if the nearest distance divided by the next closest distance is less than a certain ratio threshold, the pair of matching points is accepted. This threshold is generally between 0.4 and 0.6.

As mentioned above, the present invention converts 3D point cloud data into 2D plane data, simplifies the comparison between complex 3D point clouds into graphic matching, greatly shortens the comparison time, and the gridded image processing can greatly improve the accuracy of 3D point clouds. The smoothing requirements are greatly reduced, and it is suitable for various occasions of fast scanning and fast comparison. It is only necessary to use the same method to establish a standard library before the comparison, and the number of divisions and display colors can be changed at any time according to needs. Adding and deleting data in the database has high flexibility and strong scalability. The scale-invariant feature transformation matching algorithm ensures the rotation invariance and scale invariance of the matching, and ensures the accuracy of the comparison.

Finally, it is noted that the above embodiments are only used to illustrate the technical solutions of the present invention without limitation. Although the present invention has been described in detail with reference to the preferred embodiments, those of ordinary skill in the art should understand that the technical solutions of the present invention can be carried out Modifications or equivalent replacements without departing from the spirit and scope of the technical solution of the present invention shall be covered by the claims of the present invention.

Claims (8)

1. A method for fast search and comparison of three-dimensional point cloud planarization, characterized in that it comprises the following steps: Step 1: Obtain the 3D point cloud of the object to be measured, and then use the bilateral filter denoising algorithm to smooth the point cloud data, and then use the random sampling method to simplify the data of the smoothed point cloud data, and finally the streamlined point cloud data The point cloud data is subjected to two-dimensional transformation, and the selected view must be consistent with the view used when building the library in the standard library, and the two-dimensional point cloud image is generated after dimensionality reduction; Step 2: Find the four boundary points of the two-dimensional point cloud image through quick sorting, then generate the two-dimensional point cloud image boundary according to these four boundary points, and perform grid segmentation on the image, and then retrieve each segmented The grid, according to the density difference of the point cloud in different grids, makes corresponding marks, and fills the corresponding colors according to the marks, and generates an approximate binary image corresponding to the gradient; Step 3: Use the scale-invariant feature transformation matching algorithm to compare the obtained image with the approximate binary image in the standard library. After all the comparisons are completed, use sequential search to compare each group of images in the standard library Match the feature points of the group, find the group with the most corresponding feature points, and complete the comparison; The specific steps of the bilateral filter denoising algorithm in the step 1 are 3.1: Establish K-neighborhood Any measuring point p in the point cloud data c satisfies p∈c, then the k points closest to the measuring point p are called the K-neighborhood of p, denoted as N(p); 3.2: Normal Vector Estimation The N(p) obtained in the previous step is used to construct a plane within N(p) by the least square method, which is called the tangent plane of point p on the neighborhood N(p), denoted as T(p); T The unit normal vector of (p) is the unit normal vector at point p; 3.3: Define the viewing plane to decompose the space R ³ into the direct sum of two subspaces: , N is the one-dimensional space along the normal vector direction of the neighborhood at point p, and S2 is the ^two -dimensional tangent plane space passing through point ^p ; in the local range, define S2 as the viewing plane, and the neighborhood points are ^on the S2 plane The projection on is the position of the pixel point, and the distance from the domain point to the projection point is defined as the size of the pixel value; 3.4: Bilateral filter operator Introduce bilateral filter operator Among them, N(p) is the neighborhood of p, pi∈N(p), p' is the projection of ^p on S2, is the normal vector of point p, is the pi normal vector, W _c and _WS are Gaussian kernel functions with σ _c and σ _s as the standard deviation respectively, σ _c controls the degree of smoothness, and σ _s controls the degree of feature retention, W _c is the weight of the space domain; W _S is the weight of the feature domain; d is the distance adjusted in the direction of the normal vector, according to the formula Get smoothed coordinates . 2. A method for fast search and comparison of three-dimensional point cloud planarization according to claim 1, characterized in that: the two-dimensional transformation in the step one refers to: select the three views for generating a three-dimensional image according to the comparison needs One of the types, the view used must be the same as the view of the image in the standard library to ensure accuracy. 3. the method for a kind of three-dimensional point cloud planarization fast search comparison according to claim 1, is characterized in that: the random sampling method in the described step 1 refers to: set up a function, the random number that makes it produce can It just includes all the point clouds, and then generates a continuous set of random numbers, and then finds the corresponding points from the original point cloud and removes them until the total number of points meets the established requirements. 4. A method for fast search and comparison of three-dimensional point cloud planarization according to claim 1, characterized in that: the boundary points in the step 2 refer to the farthest points in the image, for example, in the front view In , the quick sorting method is used to find the two points with the smallest and largest X values, and the two points with the smallest and largest Y values, and a total of four coordinate points are its boundary points. 5. the method for a kind of three-dimensional point cloud planarization fast search comparison according to claim 1, is characterized in that: the grid segmentation in the described step 2 refers to: take the boundary determined in this step as a benchmark, divide The image is divided into n×n (n∈R, n>0) grid blocks. 6. The method for fast search and comparison of a kind of three-dimensional point cloud planarization according to claim 1, characterized in that: the approximate binary image in the step 2 refers to: due to the color of the image made and the color of the image in the step 2 The calibrated density mark is related, and the image made is between the grayscale image and the binary image. For the convenience of expression, the two-dimensional image used for comparison is collectively called an approximate binary image. Generate more feature points and improve the accuracy of the method. 7. A method for fast search and comparison of three-dimensional point cloud planarization according to claim 1, characterized in that: generating an approximate binary image corresponding to the gradient in the step 2 specifically comprises the following steps: traversing each grid For the point cloud in , mark each grid according to the number of point clouds in the grid, and fill the corresponding grid with the corresponding color according to the obtained corresponding mark, and the generated size is n×n(n ∈R,n>0), in order to facilitate calculation and ensure that enough feature points can be generated, two colors of black and gray are used to represent different densities, and the pointless grid is represented by white. 8. A method for fast search and comparison of three-dimensional point cloud planarization according to claim 1, characterized in that: the scale-invariant feature transformation matching algorithm in said step 3 specifically comprises the following steps: 9.1 Establishing an image scale space 9.2 Detection of key points; 9.3 Allocation of key point directions; 9.4 Feature point descriptor; 9.5 Using the exhaustive method, compare the feature points of the two images, and count the number of matching feature points for search and comparison.